Part 2: Translate Technical Claims

What they say:

“We’ll build an API integration with your CRM.”

What it means:

Two software tools talking to each other automatically — like a translator between systems.

Why you care:

• Custom API work costs $5K–$25K and takes 2–6 weeks
• Many tools already have pre-built integrations (no custom API needed)

Questions to ask:

• Is this a pre-built integration or custom API work?
• What happens if the API goes down?

What they say:

“We hit the rate limit on the integration.”

What it means:

Speed limits on how much data can move between systems — like a highway that only allows 100 cars per minute.

Why you care:

• Can break campaigns mid-flight if you send too much data at once
• Higher limits usually cost more or require enterprise plans

Questions to ask:

• What are the rate limits for our current tier?
• What happens when we hit the limit during a campaign?

What they say:

“We’ll set up a webhook to trigger the workflow.”

What it means:

An automated notification — when X happens, Y is triggered. Like a doorbell: someone presses it, you get notified.

Why you care:

• Webhooks can fail silently — you won’t know data stopped flowing
• Critical for real-time marketing automation and lead routing

Questions to ask:

• What happens if the webhook fails?
• Do we get alerts when a webhook stops working?

What they say:

“Our AI predicts customer churn with 95% accuracy.”

What it means:

AI is a broad term. Machine Learning (ML) specifically means the system learned patterns from historical data. Many vendors call basic if/then rules “AI.”

Why you care:

• Vendors commonly rebrand simple rule-based systems as “AI” to charge more
• True ML requires quality training data and ongoing maintenance

Questions to ask:

• Is this actually machine learning or is it rule-based logic?
• What data was it trained on, and how often is it retrained?

What they say:

“Our plan includes 100,000 tokens per month.”

What it means:

Chunks of text that AI processes. Roughly 750 words equals 1,000 tokens. It’s how AI vendors measure usage, like minutes on a phone plan.

Why you care:

• Overage charges range from $0.01 to $0.10 per 1,000 tokens
• A single long blog post can consume thousands of tokens

Questions to ask:

• What happens when we exceed our token limit?
• Can we monitor token usage in real time?

What they say:

“200ms latency on average for our API.”

What it means:

The delay between a request and a response — how long a user waits. 200ms is a fifth of a second.

Why you care:

• High latency breaks user experience — pages that take 3+ seconds lose 53% of visitors
• AI-powered features often add latency to existing workflows

Questions to ask:

• What’s the average response time under peak load?
• What’s the p99 latency (worst-case scenario)?

What they say:

“Our model is trained on industry-specific data.”

What it means:

The examples and information the AI learned from. Like a textbook for the AI — the quality of the textbook determines the quality of the student.

Why you care:

• Outdated training data produces outdated or wrong outputs
• Biased training data produces biased results that can harm your brand

Questions to ask:

• What data was it trained on specifically?
• When was the training data last updated?

What they say:

“We can fine-tune the model for your brand voice.”

What it means:

Further training a pre-built AI model on your specific data to customize its behavior — like giving a new employee your company handbook.

Why you care:

• Costs $5K–$50K+ depending on complexity
• Requires 1,000+ quality examples of your desired output

Questions to ask:

• How much data do we need to provide for effective fine-tuning?
• How long does fine-tuning take, and how often must we redo it?

What they say:

“Our prompt engineering ensures high-quality outputs.”

What it means:

The art and science of writing instructions for AI. Like writing a creative brief — the better the brief, the better the output.

Why you care:

• Bad prompts produce bad outputs regardless of how good the AI is
• Prompt quality is often the difference between a useful tool and a toy

Questions to ask:

• Can we customize and iterate on the prompts ourselves?
• How do you test and optimize your prompts?

What they say:

“We use a vector database for semantic search.”

What it means:

A database that stores information by meaning rather than exact keywords. Instead of searching for the word “happy,” it also finds “joyful,” “pleased,” and “delighted.”

Why you care:

• Enables genuinely smart search and recommendation features
• Can be expensive to scale — costs grow with data volume

Questions to ask:

• What’s the cost at our expected data scale?
• How does search performance change as our data grows?

What they say:

“We use a proprietary LLM for content generation.”

What it means:

The AI brain powering tools like ChatGPT, Claude, and Gemini. It’s a large neural network trained on vast amounts of text.

Why you care:

• “Proprietary LLM” is rarely true — building one costs $10M–$100M+
• Most vendors use OpenAI, Anthropic, or open-source models under the hood

Questions to ask:

• Which base model are you using (GPT, Claude, Llama, etc.)?
• What happens if that model provider changes pricing or terms?

What they say:

“We create embeddings of your content for better retrieval.”

What it means:

Converting text into numbers that capture meaning. Like creating a GPS coordinate for every piece of content so the AI can find related items by proximity.

Why you care:

• Powers semantic search, personalization, and recommendation engines
• Quality depends on the embedding model — not all are equal

Questions to ask:

• Which embedding model do you use?
• How do you handle updates when our content changes?